res an investment in new servers, optimized to process URL requests, deliver HTTP documents, serve FTP and mail, and perform other Web-related tasks.
Browsers are nearly standard equipment for any serious operating system. Whether for configuration, intranet navigation, or venturing onto the Web, the browser is evolving into an indispensable tool. And if you believe even part of the fanfare surrounding Java, your browser will soon be the interface to all applications. So what exactly does your network need to satisfy all this browsing?
Makers of high-end servers are racing to provide part of the answer. For this Lab Report, we decided to sample and evaluate symmetric multiprocessor (SMP) systems and determine how suitable they are to the emerging role of intranet servers.
The
six systems
we tested are from manufacturers that already have solid reputations in the server market: Compaq, Digital, Hewlett-Packard, Intergraph, Sun, and Silicon Graphics. During our evaluation, each of these systems proved itself quite capable. We exercised them in the role of both high-load intranet servers and in the more traditional functions of database servers.
The first three systems -- the Compaq ProLiant 5000, HP LX Pro 6/200 SMP, and Intergraph InterServe MP-660 -- are quite similar. Each is built around four Intel Pentium Pro processors in a fairly standard Intel-architecture implementation. These machines came with Microsoft Windows NT Server 4.0 installed.
The other three systems provide a challenging mix of RISC chips and different versions of Unix. The Digital AlphaServer 4100 runs OS/F, powered by four Alpha processors. The Sun Netra i 4000 server is equipped with four UltraSPARC CPUs and runs Sun's Solaris OS. Irix is the Unix of choice for SGI's quad MIPS 10000-based Origin 200 system
.
Heavy Hardware
At first it seemed natural to segregate these half-dozen servers into two categories based on their processor architecture and operating system. But as we discovered, these six machines are similar enough in performance to make such distinction artificial.
The goal of our evaluation was to rank the hardware: Specifically, we wanted to see how the multiprocessor intranet server behind the IP connection played. Typically, computers like these sit on a large corporate intranet, serving and searching a large collection of in-house documents. To handle those tasks, we specified a base server configuration built around four high-end CPUs configured for symmetric multiprocessing. Each system came equipped with 512 MB of RAM.
Because we didn't want the hard-disk subsystem to become an artificial bottleneck and bias the test results, we requested that each system come with at least 10 hard drives holding 2 to 4 GB each. We specified that they be configured as a fiv
e-drive striped data set and mirrored -- an unexceptional setup for a high-reliability server.
The unfettered flow of network traffic is obviously a key part of any intranet system. To ensure that we had more than adequate bandwidth to carry our Web activity, we specified that each machine be equipped with two 100Base-TX Ethernet ports. Clients were split evenly between the two backbones during our testing.
Beyond these minimum and necessary criteria, we allowed each manufacturer leeway to configure the system as it felt an intranet server should be equipped. This led to much customization and an interesting variety of elective equipment, as described in the features table (
note: table is not available on-line
).
Among the optional equipment that distinguished some systems were hot-swappable redundant power supplies, a built-in UPS, and a fiber-optic link. Physical form factors also varied widely, from a 6-foot-tall rack-mount system to a double-wide floor model to linked pairs of ae
rodynamic indigo towers.
As a direct result of this variation, prices for these servers ranged widely. The features table enumerates the details on each configuration; you'll have to decide if the extras justify their cost.
We didn't test systems from ALR, HP (PA-RISC), IBM (RS/6000), or NetFrame because units weren't available in time or because new models were close to introduction. Polywell supplied a system, but we were unable to get it running in time.
More Than Speed
Speed is an easily understandable metric, and it's natural to want the fastest performance possible. These quad SMP servers are, by definition, destined for heavy-traffic environments. Speed will obviously figure heavily in your decision. We found an 89 percent difference on the Web benchmark between the fastest and slowest members of our testing group. On the database test the spread was about the same, with the slowest system taking about 86 percent longer to complete the job than the fastest system.
In the real world, raw performance is weighed against costs. Bang for the buck in a pure HTML environment means serving the largest number of URL requests for the least amount of money in the shortest period of time.
A final, but by no means trivial, consideration is that of architecture and OS. Despite the generally heterogeneous nature of a typical network infrastructure, systems are not necessarily interchangeable on a practical basis.To install and manage your server, you'll need one or more experts who are familiar with the OS and architecture you choose. Developing or acquiring this expertise is often expensive.
Unless your business is a dedicated Unix shop, an NT-based server may offer a cost-effective solution for intranet serving applications. Despite its shortcomings in many areas, the general familiarity of NT Server, its growing role as a part of heterogeneous networks, and its more limited configuration options make it worth considering as a solution in this role.
Contributors
Andrew Froning,
managing editor/NSTL
Dorothy Hudson,
project manager/NSTL
Steve Platt,
director of electronic publications/NSTL
Robert L. Hummel,
freelance writer/BYTE
illustration_link (123 Kbytes)
Multiprocessor Intranet Servers
